4 - (OP14) Flexible Minds Against Extremism: A Process-based Chatbot for Public Mental Health Intervention

Modern extremism manifests in increasingly personalized and modular forms, emerging as a composite ideology dynamically constructed from various fragments of belief. Rooted in individual grievances and imbalanced motivational deficits, extremist narratives cascade through online social networks via provocative, memetic, and hedonic propaganda. The heightened vulnerability of individuals with pre-existing mental health symptoms and the self-reinforcing, rapidly spreading dynamics of extremist content online underscore the need for systematic, multilevel interventions aimed at strengthening personal and social resilience across digital and public health environments. Such an approach is not limited to targeted measures for individuals who have already adopted extremist ideologies or are at risk, but also encompasses prevention strategies aimed at enhancing public resilience, delivered through both traditional face-to-face clinic/info-house models (P/CVE framework) and digital interventions such as chatbots. This paper presents a conceptual design of a digitalized, AI-based intervention for countering extremism grounded in third-wave CBT (process-based therapy). The proposed chatbot automatically assesses users’ states through an extended evolutionary meta-model (EEMM) and suggests long-term, tailored intervention strategies by recommending adaptive processes over time to support emotion regulation and healthier social engagement. The system targets not only those already engaged in extremist ideologies but also the general public, with a preventive goal of enhancing resilience and blocking the rapid contagion of modern extremism. The chatbot supports continuous, long-term monitoring to offer tailored strategies. A RAG-based long-term memory module—using structured snapshots and episode summaries rather than a standard context window—aims to prevent failures during prolonged use and ensure consistent therapeutic strategies. The design adopts a user-centered approach that incorporates end users’ perspectives, improving acceptability for this sensitive topic while enhancing engagement and perceived rapport. Continuous safety monitoring, risk assessment, and safe-harbor de-identification are integrated to support user safety with consent-based, transparent use. A prototype is initially developed using existing chat models in a zero-shot mode, followed by a mixed-method feasibility study that recreates realistic counseling dialogues based on the EEMM framework. User feedback is applied for iterative improvement, and the resulting dialogues are annotated as seed labels for semi-supervised learning and fine-tuning to improve initial rapport, long-term personal improvement strategies, safety/risk control, and sustained engagement. Finally, a custom fine-tuned chatbot is evaluated for improvements in extremism-related outcomes [e.g., ARIS by Moskalenko & McCauley (2009) and/or clinician-rated VERA-2R], mental health (e.g., depression and anxiety), and broader individual and social well-being.